NCHLT Praat plugin
Jörg Mayer
[email protected]
April 5, 2016
Contents
1
Description
1
2
Requirements
2.1 NCHLT Speech Corpus . . . . . . . . . . . . . . . . . . . . . .
2.2 Praat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
3
3
Installation
3
4
Usage
4.1 First Usage: Corpus Base Directory . .
4.2 Language Selection and Search Pattern
4.3 Managing Results . . . . . . . . . . . .
4.4 Conclusion: The Search Cycle . . . . .
4.5 Exploit Results . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
5
9
14
14
5
Help
15
6
License
15
1
Description
This plugin enables Praat to search in the orthographic transcriptions of the
NCHLT Speech Corpus and open the audio files of corresponding search results.
The NCHLT Speech Corpus contains orthographically transcribed broadband
speech corpora for all of South Africa’s eleven official languages. Each language
corpus consists of a large training suite (trn) and a smaller test suite (tst). Launching this plugin, you first select one of your installed languages, pick one of the
suites (or both), and specify a search pattern. The XML orthographic transcription
is searched for matches of the pattern. You can further refine the search results
using speaker attributes (gender, age, location). Search results are available as a
Praat table object (including orthographic transcription, speaker ID, age, gender,
and location). Corresponding audio files can be viewed and analyzed.
2
Requirements
• Computer running Windows, OS X, or Linux
• NCHLT Speech Corpus (at least one language)
• Praat 5.4.x or newer
2.1
NCHLT Speech Corpus
At least one language of the NCHLT Speech Corpus has to be available on your
computer. The corpus may be obtained online from the South African Language
Resource Management Agency (one ZIP archive per language):
rma.nwu.ac.za/index.php/catalogsearch/result/?q=nchlt+speech+corpus
Create a new folder somewhere on your computer. You are free to choose any
name and location for this folder but be sure to remember your choice, we’ll
need this information later. This folder is called the base directory of the NCHLT
Speech Corpus in this manual. As an example, I called the base directory NCHLT
Speech Corpus and created it within a Data folder, which contains other corpora
as well:
C:
Data
NCHLT Speech Corpus
Some Other Corpus
...
Users
...
Now move the downloaded language archive(s) to the base directory and extract
the data. There should be a folder for each downloaded language inside the base
directory:
C:
Data
NCHLT Speech Corpus
nchlt_tsn
nchlt_xho
...
Some Other Corpus
...
Users
...
2
If everything went well each language folder contains two subfolders, audio and
transcriptions. Within audio you find subfolders for each speaker which contain
the actual audio files. Within transcriptions you find two XML transcription files
(training and test) and associated document type definitions:
C:
Data
NCHLT Speech Corpus
nchlt_tsn
audio
001
002
...
transcriptions
nchlt_tsn.trn.dtd
nchlt_tsn.trn.xml
nchlt_tsn.tst.dtd
nchlt_tsn.tst.xml
If you later decide to add languages just extract the new languages to the corpus base directory as decribed above. The plugin will recognize any new language.
2.2
Praat
The plugin requires Praat 5.4.x or newer. Download Praat (available for Windows,
OS X, and Linux) from: www.praat.org
You don’t need to ’install’ Praat. Just open the downloaded archive (Windows: zip,
OS X: dmg, Linux: tar.gz) and extract the Praat binary to any location on your
disk. Launch Praat at least once (double click the Praat binary) to test whether it
works well. In addition, this will create a special folder that we’ll need in the next
step.
3
Installation
Get the latest plugin version:
praatpfanne.lingphon.net/praat-ressourcen/resources-english/
Quit Praat and extract the downloaded ZIP archive (plugin_NCHLT.zip) to your
Praat preferences folder. This is a special folder which is created automatically
when Praat is launched the first time (therefore be sure to run Praat at least once
before you attempt to install the plugin). The location of this folder depends on
your operating system:
• Windows: C:\Users\YOUR-USER-NAME\Praat\
3
• OSX: /Users/YOUR-USER-NAME/Library/Preferences/Praat Prefs/
• Linux: /home/YOUR-USER-NAME/.praat-dir/
You should end up with a subfolder called plugin_NCHLT in your Praat preferences folder. Outcome on a Windows machine:
C:
Users
YOUR-USER-NAME
Praat
plugin_NCHLT
gpl.txt
manual.pdf
readme.txt
searchNCHLT.praat
setup.praat
Buttons5
Prefs5
C:\Users\YOUR-USER-NAME\Praat
plugin_NCHLT
gpl.txt
Your Praat preferences folder
The plugin subfolder
GNU General Public License
manual.pdf
This manual as PDF file
readme.txt
Concise version of the manual
searchNCHLT.praat
Main Praat script that does all the work
setup.praat Plugin setup; adds a menu entry to Praat’s Open menu and links
the main script to it
Buttons5 & Prefs5
Standard Praat preference files
That’s it. Next time, when Praat is launched you find a new entry at the bottom
of the Open menu called NCHLT: Search corpus....
4
4.1
Usage
First Usage: Corpus Base Directory
Launch Praat then select Open NCHLT: Search corpus... . The very first time you
use the plugin it will ask for the location of the corpus base directory. For this
purpose, the standard folder selection dialog of your operation system pops up
and you can navigate to the corpus base directory to select it. The corpus base
directory is the parent folder which contains all the language subfolders (see
section 2.1). In the example from section 2.1 this would be:
C:\Data\NCHLT Speech Corpus
4
Be careful to select the encompassing parent folder and not one of the language
subfolders, like e.g. nchlt_tsn.
After you selected the base directory the location is saved to a settings file, so
you don’t need to repeat this step in the future. The settings file is located in the
plugin folder and is called CorpusDir. On Windows for example:
C:\Users\YOUR-USER-NAME\Praat\plugin_NCHLT\CorpusDir
If you later decide to move your corpus to a new location or if you encounter
any problems with the plugin finding the corpus just delete that file. The plugin
will ask for the base directory the next time you run it.
4.2
Language Selection and Search Pattern
After launching the plugin, the follwing dialog box appears:
First, select one of the installed languages. If you added a new language to the
corpus since the last plugin run this language is recognized automatically and
appears in the drop-down menu. Next, select a suite. You can search in the (large)
training suite, in the (small) test suite, or in both suites at the same time. Finally,
you can specify a search pattern. Two modes are provided for the formulation of
the search pattern: simple and regex (regular expression).
5
Simple search pattern
If the simple mode is active, most of your pattern is interpreted literally, i.e. the
search results will consist of items with orthographic transcriptions that contain
exactly the specified sequence of characters. Since all NCHLT transcriptions (as
far as I know) are exclusively composed of lower-case characters, only lower-case
characters are valid in simple search patterns. The only exception to this rule is
the upper-case V character. V is a metacharacter, i.e. it is not interpreted literally
but as a placeholder for any one vowel (a, e, i, o, or u).
For example, if you are interested in the Setswana consonant cluster /tsh/ you
might select Setswana test suite (smaller but faster) and the simple search pattern
tsh:
This will produce search results like:
sa dumalane le tshwetso e
tlamela ka tshedimosetso e e
e e bontshitsweng mo setifikeiting
dikopo tsotlhe di tshwanetse go
a botshabi a tsaya dingwaga
...
6
To exclude items where tsh is followed by another consonant (e.g. w, see examples
1 and 4) use this search pattern instead: tshV, i.e. all items containing the literal
sequence tsh followed by a vowel. Results:
tlamela ka tshedimosetso e e
e e bontshitsweng mo setifikeiting
a botshabi a tsaya dingwaga
motshegang
nang le phitlhelelo go tshedimosetso
...
Of course, you can also specify literal vowels: phi searches for the literal sequence
phi. Or you can combine literal vowels and the vowel class symbol in one pattern:
phitV searches for the literal sequence phit followed by any vowel.
Regular expression search pattern
Regular expression search patterns are more complex but also much more flexible
than simple patterns. Whereas simple patterns are aware of only one metacharacter, regex patterns make extensive use of metacharacters and other special
symbols. Furthermaore, in order to compose a well-formed regular expression
of literal characters, metacharacters, and special symbols some flavor of a regex
syntax must be respected. The Praat flavor of regex syntax as well as the inventory
and semantics of metacharacters and special symbols is described in the Praat
manual:
www.fon.hum.uva.nl/praat/manual/Regular_expressions.html
Examples
In simple search mode, V is a placeholder for a set of characters, namely all
possible vowels (a, e, i, o, u). With regular expression, you can define your own
character sets. To search for tsh followed by a high front vowel, use this regex
pattern:
tsh[ie]+
tsh are literal characters matching exactly the sequence tsh. Square brackets
specify a set of characters matching either one member of the set. Hence, [ie]
matches either i or e. + is a quantifier meaning one or more occurences of the
preceding element. The complete pattern means: tsh followed by at least one of
either i or e. This finds:
tlamela ka tshedimosetso e e
e e bontshitsweng mo setifikeiting
motshegang
nang le phitlhelelo go tshedimosetso
...
7
but not:
a botshabi a tsaya dingwaga
le monontsha o o kwadisitsweng
phemithi ya mokopi wa botshabelo . . .
(BTW: The regex pattern tsh[aeiou] behaves exactly like the simple pattern
tshV.)
A range of characters in a set can be abbreviated: [a-f] = [abcdef]; all lower-case
characters: [a-z]; all digits: [0-9] etc.
The following regex pattern narrows down the previous search to word initial
occurrences of tshi or tshe:
\stsh[ie]+
\s is an escape sequence meaning whitespace, matching a space, a tabstop, begin-
ning of line etc. This finds:
tlamela ka tshedimosetso e e
nang le phitlhelelo go tshedimosetso
go naya tshedimosetso malebana le
sa lenyalo sa tshimologo kgotsa
...
but not:
e e bontshitsweng mo setifikeiting
motshegang
...
To restrict the search to phrase initial occurrences of tshi or tshe, use one of the
special anchor symbols:
ˆtsh[ie]+
The anchor ˆ matches the beginning of a line ($ matches the end of a line). This
finds only word initial occurrences at the beginning of a transcription:
tshedimosetso e e umakiwang mo
tshedimosetso nngwe le nngwe e
tshedimosetso e e leng teng
...
The last example finds occurrences of tshi or tshe in the first word of the phrase
but not necessarily in the word initial position:
ˆ\w*tsh[ie]+
\w is an escape sequence matching ’word’ characters, i.e. any letter, digit, or
underscore (this is equivalent to the character set [a-zA-Z0-9_]). The asterisk
* is a quantifier meaning zero or more occurrences of the preceding element.
So, the pattern translates to: Search at the beginning of lines looking for any
8
number (incl. zero) of word characters (i.e. letters, digits, or underscore, but not
whitespace) followed by tshi or tshe. This finds:
botshelo jo bo botoka go
motshegang
tshedimosetso e e umakiwang mo
tshelang ka bolwetsi jo bo
kgotlatshekelo o ka solofela gore
bontshitsweng mo setifikeiting sa irp
...
As mentioned above, regular expression search is complex and error prone.
A faulty regex pattern may generate results that represent only a fraction of
the data your are interested in. For this reason, be sure to carefully read the
regular expression chapter of the Praat manual before you start using the regex
mode. Additionally, you’ll find many useful online tutorials teaching regular
expressions.
When you are finished with language selection and the search pattern, press the
Search
button. The plugin loads and browses through the specified transcription, which can take some time (especially, if the large training suite was selected).
The Praat Info window gives feedback about the progress of the script—please be
patient.
4.3
Managing Results
After the search is finished, Praat info notifies you about the number of evaluated
recordings and the elapsed time. You can close that window if you wish. If the
search yields any results a new window appears: the results dialog; if nothing
was found Praat info will say so and you can try again with a different search
pattern.
9
Summary
The first section of the results dialog supplies information about search parameters and results. In the example above, the Setswana training suite (line 2) was
searched for VphV, i.e. ph between two vowels (line 3). With no filters active (line
4: gender=all, age=all, location=all). This yields 1914 recordings (line 6) from
198 speakers (line 7). Lines 8 to 10 depict speaker characteristics in terms of
gender, age range, and location: 104 of 198 speakers are female (line 8). As most
of the time, speakers with unknown age make up the largest group (115 in the
example), while speakers older than 40 are rare (5 in the example) (line 9). 140 of
198 speakers are from the Gauteng region (line 10).
Filters
The second section of the results dialog lets you refine your search by specifying
filter parameters. At the beginning, all parameters are unspecified (gender=all,
age=all, location=all).
To restrict the intitial search results to young adult speakers, select the approprite
age range (e.g. 20-30) from the drop down menu and press the REFINE button
at the bottom of the dialog. An updated results dialog will appear, summarizing
the filtered results.
10
The active filter (age range 20-30) is indicated in line 4. 399 recordings / 47
speakers remain from the intitial search results (lines 6 & 7) and all speakers are
between 20 and 30 years old (line 9).
If desired, add additional filters: Select specific gender and/or specific location
and press the REFINE button again. To reset a filter just select all from the
drop-down menu and press REFINE .
As soon as you are satisfied with your filter settings have a closer look at the
results.
Data exploration
The third section of the results dialog is devoted to data exploration. With the
TABLE
button at the bottom of the dialog the table view is toogled on and off.
Pressed the first time, a new window appears presenting symbolic information
about the results in a table. Pressed again, the table disappears.
For each found item, the table contains an index, (column "row") the path to the
associated sound file (column "audio"), the orthographic transcription (column
"orth"), and speaker characteristics (id, gender, age, location). Browse through the
table to get an overview of the data and/or to discover interesting items.
11
To open a recording for acoustic analysis, specify the recording’s index in the
recording index field of the results dialog and press the ANAYZE button. A Praat
TextGrid editor window appears with the waveform and the orthographic transcription in a TextGrid tier. When finished just close the editor window.
The recording index field is pre-filled with index 1 and automatically increased
by 1 everytime you press the ANAYZE button. This mechanism allows you to
analyze each recording one by one by just pressing ANAYZE repeatedly.
If you don’t want to analyze all recordings but only some specific ones that you
have discovered in the table, type the index of a recording into the recording
index field of the results dialog and press the ANAYZE button. You find the
recordings’ indices in the first column of the table.
For instance, if you are interested in the highlighted item below:
12
Find the index of the recording in the first dark gray column (called "row"): it’s
index 7. Type 7 into the recording index field and press ANAYZE .
When finished with the analysis, close the editor window and continue with the
next recording:
find index — type index — press
ANAYZE
Options and Exit
While the results dialog is open, you can adjust filters at any time. To return to
the original results (based on your intitial language selection and search pattern)
set all filters to all and press REFINE . Toogle the table view on and off as often
as you like.
If you analyze acoustic data consider the last option in the results dialog: remove
previous. Background: Every time you press ANALYZE , a sound file is opened by
Praat (as Sound object) and a TextGrid is generated containing the orthographic
transcription (as TextGrid object). Both objects reside in Praat’s main window
called Praat Objects. When you close the editor after the analysis, these objects
will remain in the objects list, i.e. they are still available to Praat and consume
your computer’s RAM. They’ll vanish not until you quit Praat. This is the default
behavior with remove previous unchecked. If the option is checked, as soon as
you press ANALYZE again, both previous objects (Sound and TextGrid) are
removed before the two new objects are generated. This results in a clean objects
list and better RAM hygiene. However, previous objects are removed without
warning. That’s a serious disadvantage if you plan to edit the TextGrids, because
all changes are lost when the TextGrid object is removed (unless you save the
object to a file before (!) analyzing the next recording).
Recommendation: If you don’t intend to edit TextGrids during your analysis
you can safely check this option. If you intend to edit TextGrids (adjust transcriptions, add tiers for annotations etc.) you better keep the option unchecked.
13
You have to deal with manually saving your modified TextGrids anyway and I
would recommend to do this at the latest before you close the editor window
(which would allow you to check the option). But with the option unchecked you
preserve the opportunity to catch up on saving later.
After processing all results, press EXIT at the bottom of the results dialog. The
dialog window disappears, table objects of the finished search cycle are removed,
and you can start over with a new search cycle or quit Praat.
4.4
Conclusion: The Search Cycle
• launch the plugin:
Open
NCHLT: Search corpus...
• first dialog box appears
• select language and specify search pattern; press
Search
• Praat Info appears with feedback on progress
• second dialog box (results dialog) appears
– apply filters at any time; select options from the drop-down menus
and press REFINE
– toogle table view at any time; press
TABLE
– view, analyze, and listen to audio; [type index] and press
• finish cycle; press
4.5
ANALYZE
EXIT
Exploit Results
Search and filtering results are available as Praat table objects while a search cycle
is active. The plugin’s TABLE button allows you to view these table objects,
but Praat provides many more functions to explore and analyse table objects (see
dynamic menu buttons in Praat Objects). Furthermore, it is possible to export
table objects to comma-separated files (csv format). This allows you to process
search results outside Praat.
You can find the table objects in the objects list of Praat’s main window. They are
named according to the following schema:
nchlt_language-suite-searchpattern
For instance, if you searched for VphV in the Setswana training suite the table
object is called:
nchlt_tsn-trn-VphV
(Some symbols that may be part of the search pattern—especially in regex mode—
are substituted with underscores in the context of object naming.)
14
Each search cycle generates two table objects. The first object—called like described above—contains the original unfiltered results and is not modified during
a search cycle. The second object is identical to the first object initially. As soon
as the first filter is applied ( REFINE is pressed), the suffix _refined is added to
the object name and it contains only the filtered results. Following the above
example, the second object is called:
nchlt_tsn-trn-VphV_refined
after application of a filter. If you need to process results outside Praat be sure
which table to choose (filtered or unfiltered results), select a table object accordingly and save it:
Save
Save as comma-separated file...
You have to do this before you press EXIT in the results dialog because at the
end of a search cycle both table objects are removed.
5
Help
If you find a bug or need help please send me an email:
[email protected]
6
License
Copyright © April 5, 2016
Jörg Mayer
This plugin is free software: you can redistribute it and/or modify it under the
terms of the GNU General Public License as published by the Free Software
Foundation, either version 3 of the License, or (at your option) any later version.
This plugin is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License along with
this plugin. If not, see www.gnu.org/licenses/.
15